Deep learning has revolutionized medical image analysis, enabling solutions for tasks that challenge traditional methods. This work uses deep learning to address critical bottlenecks in Adaptive Radiotherapy (ART), specifically within the complex treatment process of Total Marrow and Lymph Node Irradiation (TMLI). The research is structured around four key contributions: segmentation, image generation, fairness, and trustworthiness. First, we addressed the automation of target volume delineation for TMLI. Given the anatomical complexity and low contrast of lymph nodes in Computed Tomography (CT), we developed segmentation models to decrease the manual contouring burden of the radiation oncologist. Concurrently, we evaluated these models for algorithmic fairness, analyzing performance disparities across patient’s sex to mitigate bias. Second, we tackled data quality limitations in ART. As TMLI requires whole-body imaging, multiple Cone Beam CT (CBCT) acquisitions are aggregated together, often resulting in anatomical gaps and artifacts. We proposed a generative framework designed to inpaint these missing regions and translate the corrected CBCT volumes into high-fidelity synthetic CTs suitable for dosimetric analysis. Finally, to bridge the gap between clinical adoption and algorithmic performance, we investigated Uncertainty Quantification (UQ) and Explainable AI (XAI). We implemented methods to calibrate model confidence and visualize decision-making processes. By quantifying uncertainty and providing anatomical context for model predictions, this work aims to establish a transparent and trustworthy foundation for AI-driven radiotherapy workflows.
Il deep learning ha rivoluzionato l’analisi delle immagini mediche, abilitando soluzioni per compiti che sfidano i metodi tradizionali. Questo lavoro utilizza il deep learning per affrontare criticità fondamentali nella Radioterapia Adattativa (ART), specificamente all’interno del complesso processo di trattamento dell’Irradiazione Totale del Midollo e dei Linfonodi (TMLI). La ricerca è strutturata attorno a quattro contributi chiave: seg- mentazione, generazione di immagini, equità e affidabilità. In primo luogo, abbiamo affrontato l’automazione della delineazione del volume bersaglio per la TMLI. Data la complessità anatomica e il basso contrasto dei linfonodi nella To- mografia Computerizzata (TC), abbiamo sviluppato modelli di segmentazione per ridurre il carico di lavoro del contornamento manuale per il radioterapista oncologo. Contestual- mente, abbiamo valutato questi modelli per l’equità algoritmica, analizzando le disparità di prestazione in base al sesso del paziente per mitigare i bias. Insecondoluogo, abbiamoaffrontatolelimitazionirelativeallaqualitàdeidatinellaART. PoichélaTMLIrichiedeimmaginidell’interocorpo,vengonoaggregatediverseacquisizioni Cone Beam CT (CBCT), risultando spesso in lacune anatomiche e artefatti. Abbiamo proposto un framework generativo progettato per ricostruire queste regioni mancanti e tradurre i volumi CBCT corretti in TC sintetiche ad alta fedeltà, idonee per l’analisi dosimetrica. Infine, percolmareildivariotraadozioneclinicaeprestazionialgoritmiche, abbiamoinves- tigato la Quantificazione dell’Incertezza (UQ) e l’Intelligenza Artificiale Spiegabile (XAI). Abbiamo implementato metodi per calibrare la confidenza del modello e visualizzare i processi decisionali. Quantificando l’incertezza e fornendo un contesto anatomico per le previsioni del modello, questo lavoro mira a stabilire una base trasparente e affidabile per i flussi di lavoro radioterapici guidati dall’IA.
Computer vision applied to Adaptive Radiotherapy
Coimbra Quintas Brioso, Emanuel Ricardo
2025/2026
Abstract
Deep learning has revolutionized medical image analysis, enabling solutions for tasks that challenge traditional methods. This work uses deep learning to address critical bottlenecks in Adaptive Radiotherapy (ART), specifically within the complex treatment process of Total Marrow and Lymph Node Irradiation (TMLI). The research is structured around four key contributions: segmentation, image generation, fairness, and trustworthiness. First, we addressed the automation of target volume delineation for TMLI. Given the anatomical complexity and low contrast of lymph nodes in Computed Tomography (CT), we developed segmentation models to decrease the manual contouring burden of the radiation oncologist. Concurrently, we evaluated these models for algorithmic fairness, analyzing performance disparities across patient’s sex to mitigate bias. Second, we tackled data quality limitations in ART. As TMLI requires whole-body imaging, multiple Cone Beam CT (CBCT) acquisitions are aggregated together, often resulting in anatomical gaps and artifacts. We proposed a generative framework designed to inpaint these missing regions and translate the corrected CBCT volumes into high-fidelity synthetic CTs suitable for dosimetric analysis. Finally, to bridge the gap between clinical adoption and algorithmic performance, we investigated Uncertainty Quantification (UQ) and Explainable AI (XAI). We implemented methods to calibrate model confidence and visualize decision-making processes. By quantifying uncertainty and providing anatomical context for model predictions, this work aims to establish a transparent and trustworthy foundation for AI-driven radiotherapy workflows.| File | Dimensione | Formato | |
|---|---|---|---|
|
Ricardo_Brioso_DEIB_IT_PhD_Thesis_Politecnico_di_Milano-3.pdf
accessibile in internet per tutti
Dimensione
20.7 MB
Formato
Adobe PDF
|
20.7 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/253878